OpenAI Releases GPT-5.4, Enhancing Coding and World Knowledge for Agent Models

OpenAI has launched GPT-5.4, a new model that significantly improves coding capabilities and general world knowledge, positioning it as a strong contender for integration into agentic systems like OpenClaw. The release addresses previous limitations found in specialized coding models and offers a more balanced performance profile.

Software engineers collaborating on complex code in a modern office, symbolizing AI development.

The model's introduction comes as users seek more cost-effective and capable alternatives to existing high-performance models. GPT-5.4 is designed to bridge the gap between highly specialized coding models and those with broader world knowledge, while also offering more accessible pricing.

Performance Benchmarks and Key Improvements

GPT-5.4 demonstrates notable advancements across several benchmarks:

• GDPval: The model achieved 83.0% in GDPval, which measures performance in 44 professional knowledge work tasks, including finance and law. This score surpasses Claude Opus 4.6 (78.0%) and GPT-5.3 Codex (70.9%), indicating stronger reasoning and communication abilities in complex domains.

• SWE-Bench Pro: GPT-5.4 scored 57.7% in solving real-world software engineering problems across multiple programming languages, closely matching GPT-5.3 Codex's 56.8%. This suggests its coding proficiency remains high while other areas have improved.

• OSWorld-Verified: The model reached 75.0% in computer operation tasks, outperforming Claude Opus 4.6 (72.7%) and maintaining parity with GPT-5.3 Codex. This metric assesses an AI's ability to interact with a computer interface using mouse and keyboard commands.

• ToolAthon: GPT-5.4 achieved 54.6% in tool usage, a core indicator for agent capabilities, compared to Claude Sonnet 4.6's 44.8%.

These results indicate that GPT-5.4 combines the strong coding foundation of its predecessors with enhanced general knowledge and tool-use capabilities.

Digital interface showing performance benchmarks with graphs and charts.

Feature Updates

GPT-5.4 introduces several significant features:

• 1 Million Token Context Window: The context window has expanded to 1 million tokens, a substantial increase from GPT-5.3's 400,000 tokens. This larger context is crucial for agents, allowing them to maintain task understanding over longer interactions. OpenAI notes that usage beyond 270,000 tokens will incur double quota consumption.

• Native Computer Usage Capability: GPT-5.4 is the first mainline OpenAI model with built-in native computer usage. It can generate code to operate computers via libraries like Playwright and execute mouse and keyboard commands based on visual input. This capability is supported by a new skill, playwright-interactive, designed for debugging web and Electron applications.

• Tool Search Support: To optimize context engineering and reduce token usage, GPT-5.4 now supports tool search. Instead of receiving full tool definitions upfront, the model accesses a lightweight list of available tools and retrieves definitions only when needed. OpenAI's testing indicates this approach reduces token usage by 47% while maintaining accuracy.

Computer workstation displaying code and web application, illustrating native computer usage.

Pricing and Availability

While GPT-5.4's API pricing is higher than GPT-5.2, it remains more affordable than Claude Opus 4.6, which costs $5/$25 per million tokens for input/output. GPT-5.4 is priced at approximately half of Claude Opus 4.6.

The model is currently available on ChatGPT and Codex. While some agent platforms, such as OpenClaw, had not yet integrated GPT-5.4 at the time of its release, community demand suggests rapid adoption.

OpenAI's approach to third-party tool integration, allowing the use of subscription quotas for models like GPT-5.4, contrasts with other providers who have restricted such access or maintained higher API costs for external platforms. This policy aims to foster broader adoption and development within the agent ecosystem.